The point is that if you are the CIA, searching through a bunch of obfuscated text, just like you are already (hopefully) already compensating for basic spelling errors, it apparently would not be very difficult to expand the scope of that same mechanism (in the same way that Duck Duck Go has) to support searching through and understanding this obfuscated text.
Absolutely.
Something I've had in the back of my mind for a while is the idea of swapping out a regular Lucene tokenisation & analysis for one that treats phonemes as tokens instead of (stemmed/analysed) words being the tokens... with a similar arrangement on the query side. I think it'd have interesting capabilities... this being one.
What if you want to use it in the opposite direction? Let's call it the CIA/NSA direction, where you know what unobfuscated phrase you want to find, but not how someone may have obfuscated it. This is arguably much harder, especially if you do not store some sort of representation of the content "as sounds/phonemes".